基于 Feature-RNet 的三维大场景点云分类框架-《计算机技术与发展》

文章信息/Info

Title:: Cloud Classification Framework of 3D Large-scale Scene Based on Feature-RNet

作者:: 雷根华; 王蕾; 张志勇; 1. 东华理工大学信息工程学院,江西南昌 330013;
2. 江西省核地学数据科学与系统工程技术研究中心,江西南昌 330013

Author(s):: LEI Gen-hua1 ; WANG Lei1; 2 ; ZHANG Zhi-yong1; 1. School of Information Engineering,East China University of Technology,Nanchang 330013,China;
2. Jiangxi Engineering Technology Research Center of Nuclear Geoscience Data Science and System,Nanchang 330013,China

Keywords:: point cloud feature image; Rnet network framework; cloud classification of large-scale scene; Oakland dataset; deep learning

摘要:: 随着大场景三维点云应用在越来越多的领域中,近些年对激光点云大场景下的分类研究不断深入,各种分类模型层出不穷,在大场景点云分类任务中表现优异,但是依然存在训练时间长、计算复杂以及分类精度低等问题。针对分类精度低这一问题,提出一种基于 Feature-RNet 的三维大场景点云分类框架模型来实现点云分类工作。该框架模型不直接以三维点云数据作为输入,而是以通过 KNN 方法提取三维点云的二维特征和三维特征构建的特征图像作为输入,避免了网络框架对三维点云数据直接处理的不适应性;该模型设计的 RNet 框架结构利用了残差模块,并对其进行变型以达到提高分类精度的效果。采用公开的 Oakland 三维数据集对 Feature-RNet 框架模型进行训练,与现有的其他深度学习分类框架相比,提出的 Feature-RNet 框架模型在分类精度上有较大的提升,总体分类准确率能达到 97. 7% 。

Abstract:: With the application of large scene 3D point cloud in more and more fields,the classification research of laser point cloud inlarge scene has been deepened in recent years,and various classification models emerge in endlessly and perform well in the large-scalescenic spot cloud classification task,but there are still problems such as long training time,complex calculation and low classification accuracy. Aiming at such problem of low classification accuracy,a three - dimensional scenic spot cloud classification framework modelbased on Feature - RNet is proposed to realize point cloud classification. The framework model does not directly take the three -dimensional point cloud data as the input,but takes the two-dimensional features of the three-dimensional point cloud extracted by KNNmethod and the feature image constructed by the three-dimensional features as the input,which avoids the inadaptability of the networkframework to the direct processing of the three-dimensional point cloud data. The framework structure of RNet designed by the modeluses the residual module and modifies it to improve the classification accuracy. The open Oakland 3D data set is used to train the Feature-RNet framework model. Compared with other existing deep learning classification frameworks,the proposed Feature - RNet frameworkmodel has a great improvement in classification accuracy,and the overall classification accuracy can reach 97. 7% .